Speaker-Aware Interactive Graph Attention Network for Emotion Recognition in Conversation

Recently, Emotion Recognition in Conversation (ERC) has attracted much attention and has become a hot topic in the field of natural language processing. Conversation is conducted in chronological order; current utterance is more likely influenced by nearby utterances. At the same time, speaker dependency also plays a core role in the conversation dynamic. The combined effect of the sequence-aware information and the speaker-aware information makes the emotion's dynamic change. However, past works used simple information fusion methods to model the two kinds of information but ignored their interactive influence. Thus, we propose a novel method entitled SIGAT (Speaker-aware Interactive Graph Attention Network) to solve the problem. The core module is a mutual interactive module in which a dual-connection (self-connection and interact-connection) graph attention network is constructed. The advantage of SIGAT is modeling the speaker-aware and sequence-aware information in a unified graph and updating them simultaneously. In this way, we model the interactive influence of them and obtain the final representations, which have richer contextual clues. Experimental results on the four public datasets demonstrate that SIGAT outperforms the state-of-the-art models.